Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia

China Just Trained a Massive AI Model Without Nvidia – And It Changes Everything

For years, the conventional wisdom in artificial intelligence has been simple: if you want to train a world-class large language model, you need Nvidia GPUs. Lots of them. The company's H100 and B200 chips have become the gold standard, and the global race for AI supremacy has largely been framed as a race to buy as many of these expensive, high-demand processors as possible.

That narrative just got a lot more complicated.

Meituan, the Chinese technology giant best known for its food delivery and super-app ecosystem, has reportedly trained a massive AI model called LongCat-2.0 without using any Nvidia hardware. This is not a small experiment or a proof-of-concept run on a toy model. This is a serious, production-scale training effort that shows China can build cutting-edge AI using alternative chips — and that the global AI hardware landscape is shifting in ways that will affect every business, developer, and policymaker.

What Is LongCat-2.0 and Why Does It Matter?

LongCat-2.0 is Meituan's latest large language model, and it represents a significant leap forward in China's ability to build advanced AI systems under conditions of restricted access to American semiconductor technology. The key detail that has sent shockwaves through the tech world is that this model was trained entirely on domestic Chinese chips, with zero reliance on Nvidia GPUs.

This matters because the US government has imposed export controls that make it difficult for Chinese companies to buy the most advanced Nvidia chips. The assumption in many Western circles was that these restrictions would slow down or even halt China's progress in large-scale AI model training. LongCat-2.0 suggests the opposite may be true — or that China has found workarounds that are more effective than anyone expected.

The implications are enormous. If Chinese companies can train world-class models using domestic hardware, the entire balance of power in the AI industry shifts. The monopoly that Nvidia has enjoyed in the training market gets broken. The effectiveness of export controls as a strategic tool gets called into question. And companies around the world start wondering if they, too, can build powerful AI without depending on one supplier.

The Hardware Behind the Headline

While the exact specifications of the chips used to train LongCat-2.0 have not been fully disclosed, the general understanding is that Meituan used a combination of processors from Chinese semiconductor companies including Huawei's Ascend series and other domestic alternatives. These chips have been steadily improving over the past few years, but they have always been seen as second-tier compared to Nvidia's offerings.

That perception may need to change. Training a massive AI model requires not just raw compute power, but also sophisticated software optimizations, memory bandwidth, interconnect technology, and cooling solutions. Getting all of these to work together at scale with non-Nvidia hardware is a monumental engineering challenge. The fact that Meituan pulled it off suggests that the gap between Nvidia's chips and the alternatives is narrowing much faster than most analysts predicted.

There is also the question of software. Nvidia's dominance is not just about hardware — it is also about CUDA, the company's software platform that makes it relatively easy for developers to build and train AI models. Chinese chipmakers have been building their own software ecosystems, and LongCat-2.0 is evidence that these alternatives are now mature enough to handle large-scale production workloads.

What This Means for the Global AI Chip Market

The most immediate and obvious implication of the LongCat-2.0 story is that the global market for AI training chips is about to become much more competitive. Nvidia has enjoyed a dominant market share in this space, with competitors like AMD, Intel, and a host of startups struggling to gain a foothold. But the Chinese domestic chip ecosystem represents a massive pool of demand, innovation, and manufacturing capacity that could fundamentally reshape the industry.

Consider the numbers: China is home to some of the world's largest technology companies, including Alibaba, Tencent, Baidu, ByteDance, and Meituan itself. All of these companies are investing heavily in AI. If they can increasingly rely on domestic chips rather than Nvidia, the total addressable market for Nvidia's products could shrink significantly. This would put downward pressure on Nvidia's pricing power and margins, and it would open up opportunities for other chipmakers around the world.

It also means that the US export controls may have the opposite of their intended effect. Rather than crippling China's AI ambitions, they may have accelerated the development of a self-sufficient Chinese AI chip ecosystem. This is a classic example of the law of unintended consequences in technology policy. When you block access to a critical component, you create a powerful incentive for domestic alternatives to emerge. And when those alternatives succeed, the leverage you thought you had disappears.

The Software and Ecosystem Angle

One of the most underappreciated aspects of this story is what it says about the software stack. Training a large language model is not just about having fast chips. You need compilers, frameworks, distributed training libraries, networking protocols, and debugging tools that all work together seamlessly. Nvidia's CUDA ecosystem has been the default for years, and it has been a huge barrier for competitors.

The fact that Meituan was able to train LongCat-2.0 without Nvidia hardware means that the software ecosystem around Chinese chips has reached a new level of maturity. Developers at Meituan almost certainly had to do a lot of engineering work to make things run smoothly, but the fact that they succeeded at all is a milestone. It shows that the ecosystem is now viable for production-scale workloads.

This matters for companies outside of China as well. If you are a business in Europe, Southeast Asia, or anywhere else that is worried about relying too heavily on a single chip supplier, the Chinese alternatives are now a more credible option than they were a year ago. You may not be able to buy these chips directly due to export restrictions, but the existence of viable alternatives puts pressure on Nvidia to offer better pricing and terms everywhere.

What This Means for Businesses

For business leaders and decision-makers, the LongCat-2.0 development has several practical implications worth paying attention to.

First, the AI hardware market is no longer a one-supplier game. If you are planning a major AI infrastructure investment over the next two to three years, you should now consider a multi-supplier strategy. Relying entirely on Nvidia carries risk, and the emergence of credible alternatives gives you more bargaining power. Even if you do not switch away from Nvidia, the fact that you could switch gives you leverage in negotiations.

Second, the geopolitical risk around AI hardware is real and growing. If you are a company operating globally, you need to think about what happens if supply chains are disrupted, export controls are tightened, or tariffs are imposed. Building AI capabilities that depend on a single chip supplier is now a business risk that needs to be managed. The LongCat-2.0 story is a reminder that alternative paths exist, and that they are becoming more viable every day.

Third, the cost of AI training may come down. When there is more competition in the chip market, prices tend to fall. Domestic Chinese chips are generally less expensive than Nvidia's top-tier products, and as they become more capable, they will put pressure on the entire pricing structure. For companies that are currently spending millions of dollars on GPU clusters, this is good news. Lower training costs means more experimentation, more iteration, and faster progress.

What This Means for Society

Beyond the business implications, the LongCat-2.0 story has broader societal implications that are worth considering.

Diversification of AI hardware is good for resilience. Monocultures in technology are dangerous. If one company's chips become the backbone of the entire global AI industry, a single vulnerability — whether it is a supply chain disruption, a security flaw, or a geopolitical decision — could have catastrophic consequences. The emergence of multiple viable chip ecosystems makes the entire AI infrastructure more robust and less fragile.

It also raises important questions about regulation and safety. If Chinese companies can now train world-class AI models using domestic hardware that is outside the reach of US export controls, then the regulatory landscape becomes more complex. International coordination on AI safety standards, model evaluation, and responsible deployment becomes both more urgent and more difficult. The fact that multiple countries can now build advanced AI independently means that global cooperation is the only realistic path forward.

There is also a potential upside for innovation. When more companies and countries can participate in the cutting edge of AI research, the overall pace of progress may accelerate. The LongCat-2.0 achievement suggests that the AI revolution is not just a Silicon Valley story — it is a global phenomenon with contributions coming from all parts of the world. This diversity of perspectives and approaches could lead to breakthroughs that would not have happened in a more concentrated environment.

What Comes Next

It is important not to overinterpret a single data point. LongCat-2.0 is one model from one company. We do not yet know how its performance compares to the best models trained on Nvidia hardware. We do not know how efficient the training process was or how much it cost. We do not know if this achievement is reproducible by other companies or in other contexts.

Having said that, the symbolic significance of this event is hard to overstate. For years, the narrative has been that Nvidia is irreplaceable for large-scale AI training. That narrative is now clearly false. Chinese companies have shown that there is a path forward without Nvidia, and that path is likely to get wider and smoother over time.

In the near term, we should expect to see more announcements from Chinese companies about domestically trained models. The competitive pressure to show that you can do AI without Nvidia will drive a wave of innovation. We may also see increased investment in domestic chip startups and software tooling designed to make it easier to train models on alternative hardware.

In the longer term, the AI hardware market is likely to become more fragmented, with multiple players competing on performance, price, availability, and ecosystem quality. This is a healthy development. It means that the future of AI will not be determined by the decisions of a single company or a single country. It will be shaped by a global community of innovators working with a diverse set of tools and technologies.

Practical Takeaways for Leaders

If you are a business leader or technology executive trying to make sense of this development, here are a few concrete actions to consider.

The Bottom Line

LongCat-2.0 is more than just a technical achievement for Meituan. It is a signal that the world of AI is changing in fundamental ways. The assumption that Nvidia is the only viable platform for large-scale model training has been proven wrong. The assumption that export controls can effectively limit China's AI progress has been called into question. And the assumption that the future of AI will be controlled by a small number of Western companies and suppliers has been undermined.

None of this means that Nvidia is going away. The company still has the best hardware, the most mature software ecosystem, and the deepest relationships with the world's leading AI labs. But it does mean that the landscape is becoming more competitive, more distributed, and more interesting. For businesses, policymakers, and technologists, this is a development that demands attention — and action.

We are still in the early innings of the AI revolution, and the story of who builds the hardware that powers that revolution is far from over. LongCat-2.0 is a reminder that the future is never as predictable as it seems from the present. The best time to prepare for a multi-supplier world is now, while the transition is still underway.

TLDR: Meituan's LongCat-2.0 model, trained entirely without Nvidia chips using domestic Chinese hardware, shatters the long-held assumption that Nvidia GPUs are required for large-scale AI training. This breakthrough signals a fundamental shift in the global AI chip market, challenges the effectiveness of US export controls, and opens the door to a more competitive, multi-supplier future for AI infrastructure. Businesses and leaders should begin planning for a world where AI training no longer depends on a single chip supplier.